chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[Fix] Quick fix for int32 overflow in Mooncakes' send_kvcache_slice (#19076)

Browse Source
This commit is contained in:
YAMY
2026-02-21 20:00:33 -08:00
committed by GitHub
parent b2d8cc8cf0
commit cef353f338
1 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
python/sglang/srt/disaggregation/mooncake/conn.py
View File

@@ -460,9 +460,9 @@ class MooncakeKVManager(CommonKVManager):
)
return -1
prefill_page_indices = prefill_kv_indices.reshape(-1, 1)
decode_page_indices = dst_kv_indices.reshape(-1, 1)
tokens_per_page = np.arange(page_size, dtype=np.int32).reshape(1, -1)
prefill_page_indices = prefill_kv_indices.reshape(-1, 1).astype(np.int64)
decode_page_indices = dst_kv_indices.reshape(-1, 1).astype(np.int64)
tokens_per_page = np.arange(page_size, dtype=np.int64).reshape(1, -1)
bytes_per_token_on_prefill = src_kv_item_len // page_size
bytes_per_token_on_decode = dst_kv_item_len // page_size
src_token_slot_offsets = (